Generalization of task model using compliant movement primitives in a bimanual setting

Compliant Movement Primitives (CMPs) showed good performance for a desirable behavior of robots to maintain low trajectory error while being compliant without knowing the dynamic model of the task. This framework uses the integral representation of reference trajectories in a feedback loop together with driving joint torques that represent the feed-forward control term. To achieve CMPs generalization, refer-ence trajectories (represented in the form of task space position trajectories) are encoded as Dynamic Movement Primitives (DMPs) while the feed-forward torques are learned through the Gaussian Process Regression (GPR) and are represented as a combination of radial basis functions. This paper extends the existing framework through the generalization of CMPs in bimanual settings that can concurrently achieve low trajectory errors in relative task space and compliant behavior in absolute task space. To achieve this behavior of the bimanual robotic system, the control terms derived from CMP framework are extended with the symmetric control approach. We show how the task-specific bimanual task dynamics can be learned and generalized to different task parameters that influence the task space trajectory and to a different load. Real-world results on a bimanual Kuka LWR-4 robots configuration confirms the usability of the extended framework.